Skip to main content
info

Document Creation: [5/12/2024]. Last Edited: [10/12/2024]. Authors: [Luk Gvozdenac].

Document Code: [PE-1]. Effective Date: [10/12/2024]. Expiry Date: [10/12/2025].

Pose Estimation and Matching Project

This project focuses on pose estimation and matching using OpenCV’s DNN (Deep Neural Network) module with a pre-trained TensorFlow model. The primary goal is to detect and classify human poses in images and videos, and to compare the similarity between different poses.

Summary​

Libraries Used​

  • os
  • cv2 (OpenCV)
  • matplotlib.pyplot
  • numpy
  • pandas
  • time
  • torch

Key Components​

  1. YOLO Model Loading: The YOLO model is loaded using torch.hub.load.
  2. Pose Estimation Parameters: Parameters such as input width, height, and threshold are defined.
  3. Body Parts and Pose Pairs: Dictionaries and lists defining body parts and their connections.
  4. Image Processing: Images are read, processed, and displayed using OpenCV and Matplotlib.
  5. Data Collection: Data such as frame number, processing time, detected keypoints, and confidence scores are collected and saved to a CSV file.
  6. Pose Classification: Functions to classify poses based on detected keypoints.
  7. Video Processing: Real-time pose estimation on video feeds using OpenCV.
  8. Pose Matching: Functions to calculate the similarity between poses in two images.

Documentation​

Functions​

classify_pose(points)​

Classifies the pose based on keypoints.

yolo_pose_estimation(image)​

Performs pose estimation using the YOLO model.

pose_estimation(frame, frame_id)​

Estimates the pose in a given frame and returns the processed frame and pose label.

pose_estimation(frame)​

Estimates the pose in a given frame using OpenCV’s DNN module.

euclidean_distance(point1, point2)​

Calculates the Euclidean distance between two points.

calculate_similarity_percentage(keypoints1, keypoints2, max_distance=200)​

Calculates the similarity percentage between two sets of keypoints.

resize_to_same_height(img1, img2)​

Resizes two images to have the same height.

Data Collection​

Data is collected for each frame and saved to a CSV file for further analysis.

Real-Time Processing​

Real-time pose estimation is performed on video feeds, with the results displayed in real-time.

Pose Matching​

The similarity between poses in two images is calculated and displayed.

Usage​

  1. Image Processing: Load and process images to estimate poses and classify them.
  2. Video Processing: Perform real-time pose estimation on video feeds.
  3. Pose Matching: Compare the similarity between poses in two images.

Conclusion​

This project provides a comprehensive solution for pose estimation and matching using deep learning techniques. It can be applied to various domains such as sports analytics, fitness applications, and human-computer interaction.